change of sign of the estimates in additive models

[Diego.Pavon]

Hello,

After checking the candidates model's estimates I have realized this issue:

The model psi(), eps(Food abundance) p(Year + Location) has a NEGATIVE effect of food abundance on probability of extinction (Untransformed Estimates of coefficients for covariates (Beta's)). However, when adding one covariate to the model (psi(), eps(Food abundance + Snow Depth) p(Year + Location)) I got that food abundance as well as snow depth have POSITIVE effect on extinction. This make sense for SNOW but it does not make sense for food abundance (more food abundance, higher extinction??)

Is this a common issue in this models? Is that due to some kind of collinearity between these two variables?

Thank you

Diego

[darryl]

It sounds like potentially a collinearity issue, although the good news is that you can very easily check that.

It could also be due to sparse data so you just happen to find a model that fits the data really well. Or it could be you've found a local maximum and not the global maximum for one of these models, particularly with these reparameterized models. Check that the -2log-like value for the more complicated model is actually smaller than that for the simpler model. You might want to try different initial values and see if you come back to the same estimates.

Cheers
Darryl